fix(any): rewrite ? placeholders to $N for Postgres backend (#3000)

[allocsys]

Fixes #3000.

The Any driver always wrote ?-style placeholders, which is correct for MySQL/SQLite but not Postgres ($1, $2, ...). This caused a syntax error on any query issued through Any + Postgres that had more than a trivial shape (e.g. bound params combined with LIMIT/OFFSET).

The originally proposed fix (adding an AnyKind field to AnyArguments plus a format_placeholder override) doesn't work: QueryBuilder<Any>::push_bind() writes placeholders into the SQL string before any backend is known, since AnyArguments is built via Default::default(). The backend is only resolved later, in AnyConnectionBackend::fetch_many().

This PR instead rewrites ? -> $1, $2, ... inside sqlx-postgres/src/any.rs, right before the query is dispatched to PgConnection (covers fetch_many, fetch_optional, prepare_with, and describe). Only the Postgres leg needs this, since MySQL/SQLite already use ? natively.

The rewrite is SQL-aware (not a blind str::replace) so it skips ? that appears inside:

• single-quoted string literals (with '' escaping)
• double-quoted identifiers (with "" escaping)
• dollar-quoted blocks ($$...$$ and $tag$...$tag$)
• -- line comments and /* */ block comments

Added unit tests covering the placeholder-numbering happy path and each of the above edge cases.

CI (Examples, SQLx, SQLx CLI) passes clean on the fork: allocsys#1.

[abonander]

I was thinking that instead of trying to parse the query, QueryBuilder could just remember the positions in the string where the placeholders were inserted (because it would know) and pass that to the driver somehow.

I suppose this has the advantage of being useful even when not using QueryBuilder, but the original intent was to create our own generic placeholder (#875) which I started working on but never finished (#1076).